FPGA implementation of neuron model using piecewise nonlinear function on double-precision floating-point format

Abstract

The artificial neurons model has been implemented in a field programmable gate array (FPGA). The neuron model can be applied to learning, training of neural networks; all data types are 64 bits, and first and second-order functions is employed to approximate the sigmoid function. The constant values of the model are tuned to provide a sigmoid-like approximate function which is both continuous and continuously differentiable. All data types of the neuron are corresponding to double precision in C language. The neuron implementation is expressed in 48-stage pipeline. Assessment with an Altera Cyclone IV predicts an operating speed of 85 MHz. Simulation of 4 neurons neural network on FPGA obtained chaotic behavior. An FPGA output chaos influenced by calculation precision and characteristics of the output function. The circuit is the estimation that above 1,000 neurons can implement in Altera Cyclone IV. It shows the effectiveness of this FPGA model to have obtained the chaotic behavior where nonlinearity infuences greatly. Therefore, this model shows wide applied possibility.